CVS HEAD dumps core on simple tsvector input example
regression=# SELECT 'a very fat cat sat:4 on:5 a:6 mat:7'::tsvector;
tsvector
-----------------------------------------------
'a' 'on':5 'cat' 'fat' 'mat':7 'sat':4 'very'
(1 row)
regression=# SELECT 'a very fat cat sat:4 on:5 a:6 mat:7'::tsvector;
server closed the connection unexpectedly
Notice it's the same input both times --- only the second one crashes.
The coredump happens inside repalloc, making me suspect a memory clobber
is involved.
BTW, why does the 'a':6 lexeme disappear? To the extent that I
understand how this should work, I'd have expected 'a' and 'a':6
to merge into 'a':6 not plain 'a'.
regards, tom lane
Tom Lane wrote:
BTW, why does the 'a':6 lexeme disappear? To the extent that I
understand how this should work, I'd have expected 'a' and 'a':6
to merge into 'a':6 not plain 'a'.
'a':1,6 perhaps?
--
Alvaro Herrera http://www.amazon.com/gp/registry/DXLWNGRJD34J
We take risks not to escape from life, but to prevent life escaping from us.
Alvaro Herrera <alvherre@commandprompt.com> writes:
Tom Lane wrote:
BTW, why does the 'a':6 lexeme disappear? To the extent that I
understand how this should work, I'd have expected 'a' and 'a':6
to merge into 'a':6 not plain 'a'.
'a':1,6 perhaps?
No, it would be inappropriate to add a '1' that wasn't specified.
My reasoning is that 'a':1 and 'a':6 are distinct bits of information,
hence their combination is 'a':1,6. But 'a' doesn't give any more
information than 'a':6 so it should be dropped by the
duplicate-elimination code. It's not clear to me whether that's
what Oleg and Teodor think, though.
Hm, just found a variant of the bug:
regression=# select 'a a:6'::tsvector;
tsvector
----------
'a'
(1 row)
regression=# select 'a a:6'::tsvector;
tsvector
--------------------------------------------------------------------------
'a':6,16255C,0,12C,8C,2640,0,512,0,312,12C,400C,0,312,0,1,0,0,8448C,21,6
(1 row)
This makes it look even more like a memory-corruption issue.
regards, tom lane